Objective. Fraud detection systems operate under extreme class imbalance and strict operational alert budgets, where false positives create substantial review cost and customer friction. This study aims to develop and evaluate an explainability-aware fraud detection pipeline that is explicitly assessed at an operationally constrained decision threshold (fixed false positive rate) and to quantify how explanation patterns behave for true and false alerts in the high-precision regime.
Methods. Using the Credit Card Fraud Detection dataset (284,807 transactions; 492 frauds; 0.172% prevalence), we trained and compared four classifiers spanning interpretable baselines to high-performing ensembles: Logistic Regression, Random Forest, Histogram-Based Gradient Boosting, and XGBoost. To mirror real-world deployment, the decision threshold was tuned on a validation split to enforce FPR ≤ 1% and then applied unchanged to a held-out test set. Performance was reported using PR-AUC, ROC-AUC, recall at fixed FPR, precision, and precision@K. Statistical reliability was quantified via 1,000 stratified bootstrap resamples to obtain 95% confidence intervals. For interpretability, SHAP was used to compute global and local explanations, and we analyzed attribution concentration (fraction of total absolute SHAP mass captured by the top three features) for true positives versus false positives.
Results. At FPR ≤ 1%, all models achieved strong fraud capture (recall ≈ 0.79–0.83). Random Forest obtained the highest recall (0.829), while XGBoost achieved the highest PR-AUC (0.766). Logistic Regression remained competitive and achieved the highest precision at the operating point (0.190). Precision@100 was ~0.39–0.40 across all models, indicating comparable identification of the highest-confidence fraud cases. Explanation structure analysis showed substantial overlap between true positives and false positives, with mean attribution concentration of ~54% and ~55% respectively.
Conclusion. Under operationally fixed alert budgets, strong detection performance can be achieved while providing decision-level explanations; however, SHAP attribution patterns for false positives can closely resemble those for true fraud in high-precision regimes. Explanations are valuable for transparency and investigator context, but they should be treated as decision support rather than a reliable discriminator for triage on their own.If you have any questions about submitting your review, please email us at [email protected].